AI Coding Agents Write Their Own Report Cards — and That's the Core Problem

A developer building a project with Claude Code discovered on July 27 that his task tracker had drifted badly from reality: a card held a null commit for two days, the spec contained nine false statements, and 500 lines of code were written against a task never formally started. The root cause was structural — the AI agent both performed the work and self-reported its status, with no independent verification layer in between. Standard fixes like stricter prompt files and better hooks provided marginal improvement but did not address this fundamental flaw. The developer ultimately redesigned his workflow so that status changes require machine-verified checks defined by humans in project config, while Git itself arbitrates whether work actually landed rather than relying on the agent's claims. The key insight is that verification authority — deciding which surface can assert a fact — matters more than adding more rules for the agent to follow.